Electric Energy Demand Forecasting with Explainable Time-series Modeling

Recently, deep learning models are utilized to predict the energy consumption. However, to construct the smart grid systems, the conventional methods have limitation on explanatory power or require manual analysis. To overcome it, in this paper, we present a novel deep learning model that can infer the predicted results by calculating the correlation between the latent variables and output as well as forecast the future consumption in high performance. The proposed model is composed of 1) a main encoder that models the past energy demand, 2) a sub encoder that models electric information except global active power as the latent variable in two dimensions, 3) a predictor that maps the future demand from the concatenation of the latent variables extracted from each encoder, and 4) an explainer that provides the most significant electric information. Several experiments on a household electric energy demand dataset show that the proposed model not only has better performance than the conventional models, but also provides the ability to explain the results by analyzing the correlation of inputs, latent variables, and energy demand predicted in the form of time-series.